Relevant Algorithms For General Object Recognition Based On Feature Combination

General Object Recognition (GOR) is a basic but hard problem in Vision. Althoughthere exist a number of working algorithms and systems for specific purpose recognition task(e.g., Face Detection & Recognition), to develop a general-purpose recognition algorithmand system with human-level performance, is still challenging.Vision Research is interdisciplinary, covering a variety of fields from Computer Visionto Neurobiology. This paper firstly makes a survey on recent achievements of GOR, with anemphasis on modeling techniques and computational approaches in the regime of ComputerVision. In addition, one primary approach for GOR, the patch-based method, is furtherstudied.Then the famous biologically plausible HMAX model is elaborated and analyzed inthis thesis. The hierarchical structures of HMAX is fully decomposed and its functionalcounterpart in Computer Vision is clearly revealed. Based on the insights, two novel localimprovement strategies are developed and the experiment results show that both are indeedpractical.The main contributions of this thesis are as follows:1) This paper provides an elaborate description on computational feasible HMAX model.Four-layered structures (S1,C1,S2,C2) of HMAX model are under close scrutiny. Asa result, its functional relation to the counterparts in Computer Vision is revealed:"blocky histogram"(S1+C1) by"histogram"(S2+C2). Due to this analysis, HMAXmodel can be called a supervised"one layer and a half"model in term of featurecombination, if we take one stage of"blocky histogram"extraction as one layer. 2) Some tentative improvement strategies are developed in local operations of HMAXmodel. Different from regular strategies that focus purely on machine learning, in thispaper two novel vision-based approaches are proposed: a) A Non-uniform partition onblocky histogram in contrast with fixed 8×8 partitions in HMAX, and b) A globalscheme to find common parts of multiple images, in contrast with local and random-ized patch-picking in HMAX.3) Practical modeling and solutions of both proposed strategies are given. For a), its 1-Dand 2-D case are analyzed respectively. For 1-D case('signal case'), it can be solvedexactly and efficiently via Dynamic Programming, while for 2-D case ("image case"),no exact solution is available. Instead, coordinate descent is adopted to get a quitereasonable approximation. For b), a generative model is proposed and standard EMiterations are used to get locally best solutions of the problem. For both strategies,experiments are conducted. The experiment results prove that both ideas lead to betterperformance and thus are applicable.4) This paper constructs a patch-based interactive platform for GOR. It is designed tofacilitate algorithm debugging in Computer Vision. On this platform, users can see theresult of feature extraction and the running process of algorithms. They can dynami-cally adjust parameters as well. Besides, this platform also has some novel character-istics on software architectures.